Performance Comparison of Clipping

Detection Techniques in AC Power

Time Series
Preprint
Kirsten Perry, Matthew Muller, and Kevin Anderson
National Renewable Energy Laboratory
Presented at the 48th IEEE Photovoltaic Specialists Conference (PVSC 48)
June 20-25, 2021

NREL is a national laboratory of the U.S. Department of Energy Conference Paper

Office of Energy Efficiency & Renewable Energy NREL/CP-5K00-78954
Operated by the Alliance for Sustainable Energy, LLC October 2021
This report is available at no cost from the National Renewable Energy
Laboratory (NREL) at www.nrel.gov/publications.

Contract No. DE-AC36-08GO28308

 Performance Comparison of Clipping
Detection Techniques in AC Power
Time Series
Preprint
Kirsten Perry, Matthew Muller, and Kevin Anderson
National Renewable Energy Laboratory

Suggested Citation
Perry, Kirsten, Matthew Muller, and Kevin Anderson. 2021. Performance Comparison of
Clipping Detection Techniques in AC Power Time Series: Preprint. Golden, CO: National
Renewable Energy Laboratory. NREL/CP-5K00-78954.
https://www.nrel.gov/docs/fy22osti/78954.pdf.

© 2022 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in
any current or future media, including reprinting/republishing this material for advertising or promotional purposes,
creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of
this work in other works.

NREL is a national laboratory of the U.S. Department of Energy Conference Paper

Office of Energy Efficiency & Renewable Energy NREL/CP-5K00-78954
Operated by the Alliance for Sustainable Energy, LLC October 2021
This report is available at no cost from the National Renewable Energy National Renewable Energy Laboratory
Laboratory (NREL) at www.nrel.gov/publications. 15013 Denver West Parkway
Golden, CO 80401
Contract No. DE-AC36-08GO28308 303-275-3000 • www.nrel.gov
 NOTICE

This work was authored by the National Renewable Energy Laboratory, operated by Alliance for Sustainable
Energy, LLC, for the U.S. Department of Energy (DOE) under Contract No. DE-AC36-08GO28308. Funding
provided by the U.S. Department of Energy’s Office of Energy Efficiency and Renewable Energy (EERE) under
Solar Energy Technologies Office (SETO) Agreement Numbers 34348. The views expressed herein do not
necessarily represent the views of the DOE or the U.S. Government. The U.S. Government retains and the
publisher, by accepting the article for publication, acknowledges that the U.S. Government retains a nonexclusive,
paid-up, irrevocable, worldwide license to publish or reproduce the published form of this work, or allow others to
do so, for U.S. Government purposes.

This report is available at no cost from the National Renewable

Energy Laboratory (NREL) at www.nrel.gov/publications.

U.S. Department of Energy (DOE) reports produced after 1991

and a growing number of pre-1991 documents are available
free via www.OSTI.gov.

Cover Photos by Dennis Schroeder: (clockwise, left to right) NREL 51934, NREL 45897, NREL 42160, NREL 45891, NREL 48097,
NREL 46526.

NREL prints on paper that contains recycled content.

 Performance Comparison of Clipping Detection
Techniques in AC Power Time Series
Kirsten Perry1 , Matthew Muller1 , Kevin Anderson1
1
NREL, Golden, CO, USA

Abstract—In this research, a variety of methods were devel-

oped to detect clipping periods in AC power time series. AC
power data streams associated with 36 unique systems across
the United States were collected, and data points representing
clipping periods were manually labeled by experts. Using this
data set for training and validation, novel logic-based and
machine learning (ML) approaches were developed to classify
time series values as clipping or non-clipping. These approaches
were compared to the RdTools method for detecting clipping
periods. The logic-based and ML XGBoost approaches achieved
F-scores of 85.0 and 77.6, respectively, when cross-validated
against the manually labeled data, as compared to the current
RdTools approach (F-score of 56.4), indicating a significant
improvement at detecting clipping periods. Additionally, the
effects of each clipping filter when evaluating system degradation
rates were assessed, using 31 unique systems across the United
States. Results indicate that estimated system degradation rate
can vary based on the type of clipping filter used, by up to 0.6%
degradation rate for some cases.
Index Terms—machine learning, clipping, photovoltaic, solar,
modeling, rdtools

I. I NTRODUCTION
Due to a significant decrease in solar module prices over the
past decade, developers have increased installations’ inverter
loading ratio (ILR), resulting in a higher DC-to-AC ratio.
By increasing a system’s DC-to-AC ratio, developers can
increase system output outside of peak irradiance windows.
Consequently, this design decision has led to an increase in
clipping in solar projects. Clipping prevents inverter overload
by operating the array in a reduced efficiency state when nor-
mal operation would exceed the inverter’s power conversion
limits. Generally, clipping signals manifest as a flat line at or
near the peak of an AC inverter’s production capacity. Clipping Fig. 1. Example AC power time series data, taken from actual field PV
systems. Clipping periods are labeled in yellow and non-clipping periods are
thresholds can remain constant for an AC power time series, labeled in blue.
or may vary. Some causes of variance include temperature
derating and dynamic plant control. Some example clipping
profiles, taken from physical AC power streams, are shown in analysis. Although this approach is simple and intuitive, it fails
Fig. 1. to handle edge cases where the clipping cutoff threshold varies
Best practices for PV degradation analysis include the over the course of the time series. Furthermore, this approach
removal of clipped points from AC power or energy time series filters out time series data during periods of peak irradiance,
[1]. A few approaches have been explored to mask clipping pe- where un-clipped systems may be performing optimally.
riods in AC inverter time series, which are described presently. Another approach developed for clipping detection uses
One approach, given in the RdTools Python package [2], fuzzy system logic [3], where features such as actual plant
is to filter power or energy data using a percentile cutoff [1]. power, the time series gradient mean, and the time series
The default clipping filter in the RdTools package is set to gradient range are taken as inputs to evaluate the degree of
the 98th quantile, where any data in the AC power time series clipping in a solar plant. This method was validated using
greater than 99% of the 98th quantile is assumed as clipping, labeled data from a PV plant located in Spain, and was
and filtered out of the time series before running a degradation approximately 90% accurate. This research only explored

1
This report is available at no cost from the National Renewable Energy Laboratory at www.nrel.gov/publications.
algorithm performance at a single location; in contrast, our TABLE I
work handles multiple configurations, data frequencies, and D ISTRIBUTION OF M ANUALLY LABELED DATA , BY AVERAGING I NTERVAL
locations, so it is more universally applicable when applied to
field data. Frequency (Minutes) Number Data Streams Number Readings
In this paper, we present a series of logic-based and ML 1 15 951387
15 18 1048464
clipping filters and compare their performance to the RdTools 30 1 64232
clipping filter performance using manually labeled and simu- 60 2 61364
lated clipping validation data. We plan to publicly release these
algorithms via the RdTools package, as well as the manually
labeled data that we used during the validation process. TABLE II
D ISTRIBUTION OF M ANUALLY LABELED DATA , BY M OUNTING
Additionally, we assess the effects of applying each clipping C ONFIGURATION
filter when performing system degradation analysis, using
physical PV system data. By doing this, we demonstrate that Mounting Configuration Number Data Streams Number Readings
degradation estimates vary by clipping filter used. Fixed Tilt 21 1237417
Single-Axis Tracking 15 888030
II. M ETHODOLOGY
A. Datasets
Thirty-six AC power time series, representing 31 physical at 1-minute, 15-minute, and 30-minute intervals. Seventeen
PV systems and approximately 2.13 million rows of data, were of the test cases are associated with a fixed-tilt mounting
selected with the intent of building a training and validation configuration, and six of the test cases are associated with
set. Data with a wide range of clipping behaviors was selected, a single-axis tracking mounting configuration.
including varying clipping signal behavior (noisy vs. clean
signals), averaging interval (ranging from 1-minute to hourly), B. RdTools Approach
and mounting configuration (single-axis tracking vs. fixed
Each AC power time series described in the Datasets
tilt). Estimated clipping periods for each time series were
subsection was run through the RdTools clipping filter. This
manually labeled by two experts with experience identifying
filter calculates the 98th quantile of the AC power time series,
clipping signals in data. The datasets used for this study were
and any values that are greater than 99% of the 98th quantile
provided via the National Renewable Energy Laboratory’s
are classified as clipping.
PV Fleet Performance Data Initiative [4]. A summary of
Any clipping periods identified by the RdTools clipping
the manually labeled data set distribution by data averaging
filter were compared to the labeled training sets, to quantify
interval and mounting configuration is provided in Tables I and
filter performance. Filter accuracy, precision, recall, and F-
II, respectively. Predominantly, data streams are averaged at 1-
score were calculated using the following three equations,
minute and 15-minute intervals, with one data stream averaged
respectively:
at a 30-minute interval and two data streams averaged at a
60-minute interval. Fifteen of the data streams are associated TP + TN
with single-axis tracking systems, and twenty-one of the data accuracy = (1)
TP + FP + TN + FN
streams are associated with fixed tilt systems.
A subset of the manually labeled data was selected and TP
artificial clipping signals were inserted into these time se- precision = (2)
TP + FP
ries. This simulated clipping data set was generated to test
algorithm performance during ideal clipping situations, where TP
all data are clipped at a constant AC power threshold. Two recall = (3)
TP + FN
different approaches were taken for inserting artificial clipping
signals. First, some data sets were down-sampled at different 2 · recall · precision
frequencies, and time periods labeled as 100 percent clipping F-score = (4)
recall + precision
were marked as clipping periods. For the second approach,
clipping at a variable threshold above the 70th percentile was where T P is the number of true positives, F P is the number
inserted in the time series, by setting all values greater than or of false positives, F N is the number of false negatives, and
equal to the threshold as the associated threshold value. Each T N is the number of true negatives. Accuracy represents the
time series could have up to 5 different clipping thresholds number of correctly classified data points divided by total data
across the series, with location and number of thresholds points. Recall, or true positive rate (TPR), is the ability to
randomly selected. A summary of the simulated clipping data, correctly detect positive cases. Precision, or positive prediction
broken down by averaging interval and mounting configura- value (PPV), is the fraction of correctly identified positives
tion, is provided in Tables III and IV, respectively. Overall, the given the total positive predictions. F-score is the harmonic
simulated clipping data set consists of twenty-three individual mean of precision and recall. The higher the F-score, the more
test cases, with the vast majority of the data sets averaged precise and robust the classifier.

2
This report is available at no cost from the National Renewable Energy Laboratory at www.nrel.gov/publications.
 TABLE III values is determined. A maximum and minimum daily
D ISTRIBUTION OF DATA WITH S IMULATED C LIPPING P ERIODS , BY threshold for setting clipping values is derived as 2 stan-
AVERAGING I NTERVAL
dard deviations above and below the mean, respectively.
Frequency (Minutes) Number Data Streams Number Readings
For time series data sets with a averaging interval less
frequent than every 10 minutes, the daily maximum and
1 5 484783
5 1 199993 minimum thresholds for clipping are set as the maximum
15 9 390142 and minimum values where clipping is detected over the
30 6 203501 course of that day, respectively.
60 2 61322
6) Once the daily maximum and minimum clipping thresh-
olds are determined, all data values between these two
TABLE IV thresholds are updated as clipping periods.
D ISTRIBUTION OF DATA WITH S IMULATED C LIPPING P ERIODS , BY 7) An overall clipping threshold is applied to the data,
M OUNTING C ONFIGURATION
based on the classified clipping periods derived in the
Mounting Configuration Number Data Streams Number Readings
previous steps. The following is calculated:
Fixed Tilt 17 1162365 xp99 − clippedxp99
thresholdclipping = xp99 −clippedxp99
(6)
Single-Axis Tracking 6 177376
2
In addition to the previously identified clipping points,
all values in the time series greater than the overall
C. Logic-Based Approach
clipping threshold are classified as clipping.
A new logic-based filter was developed for detecting clip- 8) Accuracy, precision, recall, and F-score, described in
ping periods in time series. This filter identifies clipping (1)-(4), are calculated, where the logic-based model
periods via the following steps: predictions are compared to the ground truth labels in
1) The averaging interval of the power or energy time series the training and validation data set.
is determined. If the averaging interval is more frequent
than once every 10 minutes, then the time series is D. Supervised Machine Learning Approach
aggregated to 15-minute intervals, using the mean. The labeled time series data described in the Datasets
2) The maximum and minimum readings are calculated section are used to train a series of black-box ML models
across a moving window for the time series. For tracked to classify clipping periods. For each individual data stream,
systems with a data averaging interval more frequent the data is min-max normalized, via the following equation:
than once every 30 minutes, a window of 5 readings is x − min(x)
used. For all other systems, a window of 3 readings is xscaled = (7)
max(x) − min(x)
used.
Once the maximum and minimum moving values are By performing min-max normalization, data streams with
determined, the following equation is used to calculate different units and clipping thresholds can be compared si-
the maximum rolling range of each value: multaneously. Features are then derived from the min-max
normalized data. Features used in the model include:
xrolling max − xrolling min
rangemax = xrolling max +xrolling min ∗ 100 (5) • The min-max normalized AC power time series
2 • The simple moving average of the time series, where a
where xrolling max is the rolling maximum over the past 5 reading-window is used when the averaging interval is
n values, and xrolling min is the rolling minimum over the more frequent than once every 10 minutes, a 3 reading-
past n values. window is used when the averaging interval is between
3) An initial boolean clipping mask is derived, where 10 minutes and an hour, and 2-reading window is used
values with a max rolling range (see (5)) of less than 0.2 otherwise
are marked as clipping. When an individual data point is • The forward- and backward-calculated derivatives for the
determined as clipped per (5), all data points within the min-max normalized values and the rolling average
given p-length rolling window are also set as clipped, • The maximum rolling range of the min-max normalized
where p is 3 or 5 reading, as described in Step 2. The values, as shown in Equation 5
value 0.2 is an empirically-derived cutoff. • The mounting configuration (fixed tilt vs. single axis
4) Any high-frequency time series that were aggregated up tracking), converted to a categorical variable
to 15-minute intervals are sampled back to their original • The mode, or most common, time series sampling fre-
frequency. Clipping labels are forward-filled in the data. quency
5) Additional logic is added to address noise in high • The daily maximum of the time series, as well as the
frequency time series sets. For time series data sets with percent of the daily max each individual reading is
a averaging interval of 10 minutes or more frequent, the • After generating classifier predictions for a data set,
daily mean and standard deviation for labeled clipping additional override logic is introduced. Specifically, all

3
This report is available at no cost from the National Renewable Energy Laboratory at www.nrel.gov/publications.
 values between the maximum and minimum predicted TABLE V
clipping periods for each day are set as clipping, where C LIPPING F ILTER P ERFORMANCE ON LABELED C LIPPING DATA
predicted clipping periods must be 80 percent or greater
Label Algorithm Accuracy(%) Recall Precision F-Score
of the associated daily maximum value.
Manual RdTools 92.1 46.1 72.6 56.4
A random forest classifier model and an extreme gradient Manual Logic-based 96.5 89.4 81.0 85.0
boosting, or XGBoost, classifier model were used to train Manual Random Forest 95.2 79.4 77.6 78.5
Manual XGBoost 94.8 81.4 74.1 77.6
the model. Python’s open-source scikit-learn [5] package was Simulated RdTools 98.0 89.1 96.9 92.8
used to build the random forest model, and the open-source Simulated Logic-based 98.1 95.6 89.5 92.4
XGBoost package [6] was used to build the XGBoost model. Simulated Random Forest 98.5 90.9 96.5 93.6
Simulated XGBoost 97.3 81.8 95.4 88.1
Cross-validation was used when evaluating model perfor-
mance. For each individual data stream, the model was re-
trained with the data stream omitted from the training set,
and model performance was evaluated on the omitted data RdTools filter is constructed specifically to detect. However,
stream. In all, the model was regenerated 59 times, for each under field conditions, clipping signals do not necessarily
of the 59 data streams (manually labeled and simulated data). manifest this ideal behavior, as the clipping threshold may
After the model had generated clipping classifications for each vary over time and include noise. All of the algorithms perform
data stream, the overall accuracy, precision, recall, and F-score better on the simulated data than the manually labeled data,
were calculated. likely because simulated data represents idealized clipping
behavior, in opposition to the real-world behavior present in
E. Applying Algorithms to Field Data the manually labeled data set.
The degradation rates of 31 different systems in the PV Table VI illustrates algorithm performance by mounting
Fleets Initiative were calculated using the RdTools sensor- configuration (fixed tilt vs. single-axis tracking), and Table VII
based methodology [4]. Systems comprised a series of dif- illustrates algorithm performance by data averaging interval,
ferent solar fleet owners, as well as diverse geographic loca- for each of the data sets. In examining the results from
tions, sampling frequencies, and system configurations. When Table VI, algorithms generally perform similarly for fixed-
performing the degradation analysis, all parameters remained tilt systems and single-axis tracking systems, when F-score is
constant, except the clipping filter used on the data. Five differ- used as a benchmark. One notable exception is the RdTools
ent experiments were run to assess the role the clipping filter filter, which performs significantly better on fixed tilt data set
has on degradation results: no clipping filter applied to the compared to single-axis tracking data set, for the manually
data (the placebo); the RdTools filter applied; the logic-based labeled data (an F-score of 67.7 vs. an F-score of 51.6,
algorithm applied; the Random Forest algorithm applied; and respectively). AC power data for a tracking system manifests
the XGBoost algorithm applied. Once the RdTools sensor- as a valley-like signal around maximum daily output, as shown
based degradation methodology was run on the systems, the in Figure 2. Although this signal is distinct from a clipping
results for each experiment were compared. signal, it is likely to be confounded as a clipping signal using
the RdTools’ filter’s simple logic.
III. R ESULTS
Algorithm performance varied across averaging interval
A. Algorithm Performance on the Training and Validation Set for the manually labeled data set. For 1-minute, 15-minute,
Table V displays the accuracy, precision, recall, and F- and 60-minute manually labeled data, the logic-based filter
score for each of the different clipping algorithms, based on had the best performance, with F-scores of 81.1, 90.1, and
performance on the manually-labeled data set and simulated 90.3, respectively. The random forest algorithm had the best
clipping data set, respectively. performance on the 30-minute manually labeled data, with an
Using F-score as a benchmark since it accounts for both F-score of 90.6.
classifier precision and robustness, the best overall perfor-
mance on the manually labeled data set was achieved by the B. Degradation Variability for Physical Systems, based on
logic-based clipping filter, followed by the Random Forest Clipping Algorithm Used
and XGBoost clipping filters, respectively. All ML and logic- In total, 31 systems, representing multiple fleet owners
based filters performed better on the manually labeled data set, across the NREL PV Fleets initiative, were evaluated using
compared to the standard RdTools clipping filter (F-score of the RdTools methodology to determine sensor-based degra-
56.4). For the simulated data set, the Random Forest classifier dation rates [4]. Systems represented a wide geographic area
algorithm had the best overall performance with an F-score across the United States, with multiple configurations, module
of 93.6, followed by the RdTools filter (F-score of 92.8) technologies, and data sampling frequencies. Particularly, we
and the logic-based filter (F-score of 92.4), respectively. It is focused on how estimated degradation rates varied for each
unsurprising that the RdTools filter scores significantly better system, based on which clipping filter was used. Fig. 3 shows
on simulated data than on physical data. The simulated data the difference in sensor-based degradation rates, where the
sets focus largely on ideal conditions where clipping occurs at results that used the logic-based, XGBoost, and RdTools filters
a consistent threshold at the time series maximum, which the are subtracted from the results where no filter was used, for

4
This report is available at no cost from the National Renewable Energy Laboratory at www.nrel.gov/publications.
 TABLE VI
C LIPPING F ILTER P ERFORMANCE B Y M OUNTING C ONFIGURATION

Labeling Mounting Algorithm Accuracy (%) F-Score

Manual Fixed Tilt RdTools 97.0 67.7
Manual Fixed Tilt Logic-based 98.3 82.2
Manual Fixed Tilt Random Forest 97.9 76.6
Manual Fixed Tilt XGBoost 97.8 76.9
Manual Single-Axis Tracking RdTools 85.3 51.6
Manual Single-Axis Tracking Logic-based 93.7 83.1
Manual Single-Axis Tracking Random Forest 91.5 79.1
Manual Single-Axis Tracking XGBoost 90.6 77.9
Simulated Fixed Tilt RdTools 98.1 92.6
Simulated Fixed Tilt Logic-based 98.1 90.8
Simulated Fixed Tilt Random Forest 98.6 92.2
Simulated Fixed Tilt XGBoost 97.2 83.7
Simulated Single-Axis Tracking RdTools 96.6 94.0
Simulated Single-Axis Tracking Logic-based 98.2 96.6
Simulated Single-Axis Tracking Random Forest 98.2 96.7
Simulated Single-Axis Tracking XGBoost 98.2 96.8

TABLE VII
C LIPPING F ILTER P ERFORMANCE B Y AVERAGING I NTERVAL

Labeling Frequency (min) Algorithm Accuracy (%) F-Score

Manual 1 RdTools 85.3 49.8
Manual 1 Logic-based 92.9 81.1
Manual 1 Random Forest 90.6 76.7
Manual 1 XGBoost 89.7 75.5
Manual 15 RdTools 97.8 76.4
Manual 15 Logic-based 99.3 90.1
Manual 15 Random Forest 98.9 85.8
Manual 15 XGBoost 99.0 86.8
Manual 30 RdTools 98.7 63.2
Manual 30 Logic-based 99.4 73.1
Manual 30 Random Forest 99.8 90.6
Manual 30 XGBoost 99.7 89.6
Manual 60 RdTools 94.0 43.5
Manual 60 Logic-based 98.7 90.3
Fig. 2. Tracking system behavior vs. fixed tilt system behavior, respectively, Manual 60 Random Forest 98.5 88.4
where clipping periods are labeled in yellow and non-clipping periods are Manual 60 XGBoost 98.3 87.2
labeled in blue. Simulated 1 RdTools 97.5 93.2
Simulated 1 Logic-based 96.8 92.0
Simulated 1 Random Forest 98.8 96.9
Simulated 1 XGBoost 96.1 89.7
each respective system. Data is color-coded by system, and Simulated 15 RdTools 97.1 88.7
system degradation rate is calculated as the mean degradation Simulated 15 Logic-based 99.5 98.2
across all AC power inverters associated with the system. Simulated 15 Random Forest 96.7 81.7
Simulated 15 XGBoost 96.0 77.1
Based on these results, the type of filter affects degradation Simulated 30 RdTools 99.3 96.8
estimates; in some systems, this can be up to 0.6%/year. Simulated 30 Logic-based 99.7 98.5
Simulated 30 Random Forest 99.7 94.7
Figure 4 displays the individual distributions of the clipping Simulated 30 XGBoost 99.5 91.6
filter variation, where the system degradation rate with each Simulated 60 RdTools 91.3 76.5
Simulated 60 Logic-based 99.2 98.1
respective filter applied is subtracted from the system degra- Simulated 60 Random Forest 99.5 98.8
dation rate when no filter is applied. The median difference Simulated 60 XGBoost 99.4 98.6
in degradation estimates for the XGBoost, Random Forest,
Logic-Based, and RdTools filters was 0.05%, 0.05%, 0.04%,
and 0.06%, respectively. Likewise, the standard deviation of known degradation rates, to evaluate degradation accuracy
the differences for the XGBoost, Random Forest, Logic- when each of the four treatment types was used. By doing this,
Based, and RdTools filters was 0.11%, 0.12%, 0.10%, and we can further demonstrate that improving clipping filtering on
0.20%, respectively. These results indicate that the degradation AC power data streams results in more accurate degradation
distributions for the XGBoost, Random Forest, and logic-based estimates for a system.
filters are similar, with a tighter distribution than the RdTools
degradation results. V. C ONCLUSIONS
This research presents a series of new filters that can be
IV. F UTURE W ORK used to successfully detect clipping periods in AC power time
In addition to performing degradation analysis on physical series data from solar installations. These filters, which will be
systems, we plan to the same process on simulated data with released in the RdTools Python package, allow solar analysts

5
This report is available at no cost from the National Renewable Energy Laboratory at www.nrel.gov/publications.
 are used to mask AC power time series data. In our analysis of
31 physical systems, we show that sensor-based degradation
rates vary based on the clipping filter used on a system-by-
system basis, in some cases by up to 0.6%. Additionally,
system results show that the degradation distributions for
the newly developed XGBoost, Random Forest, and logic-
based filters are similar, with a tighter distribution than the
RdTools degradation results. We plan to further investigate if
improving clipping filter performance leads to more accurate
system degradation results, using simulated data with known
degradation rates.
ACKNOWLEDGMENT
This work was authored by Alliance for Sustainable Energy,
LLC, the manager and operator of the National Renewable
Energy Laboratory for the U.S. Department of Energy (DOE)
under Contract No. DE-AC36-08GO28308. Funding provided
by the U.S. Department of Energy’s Office of Energy Effi-
ciency and Renewable Energy (EERE) under Solar Energy
Technologies Office (SETO) Agreement Numbers 34348. The
views expressed in the article do not necessarily represent
the views of the DOE or the U.S. Government. The U.S.
Fig. 3. Scatterplot of the difference in sensor-based degradation rates when no Government retains and the publisher, by accepting the ar-
filter is used vs. when each respective filter is used, color-coded by individual ticle for publication, acknowledges that the U.S. Government
system
retains a nonexclusive, paid-up, irrevocable, worldwide license
to publish or reproduce the published form of this work, or
allow others to do so, for U.S. Government purposes.
R EFERENCES
[1] D. C. Jordan, C. Deline, S. R. Kurtz, G. M. Kimball, and M. Anderson,
“Robust PV degradation methodology and application,” IEEE Journal
of Photovoltaics, vol. 8, no. 2, pp. 525–531, Mar. 2018. [Online].
Available: https://doi.org/10.1109/jphotov.2017.2779779
[2] M. Deceglie et al., “NREL/rdtools: Version 2.1.0-beta.1,” 2020. [Online].
Available: https://doi.org/10.5281/ZENODO.4307010
[3] F. S. Fernández, M. A. Muñoz-Garcı́a, and S. Saminger-Platz, “Detecting
clipping in photovoltaic solar plants using fuzzy systems on the feature
space,” Solar Energy, vol. 132, pp. 345–356, Jul. 2016. [Online].
Available: https://doi.org/10.1016/j.solener.2016.03.013
[4] C. Deline, R. White, M. Muller, K. Anderson, K. Perry, M. Deceglie,
L. Simpson, and D. Jordan, “PV fleet performance data initiative
program and methodology,” in 2020 47th IEEE Photovoltaic Specialists
Conference (PVSC). IEEE, Jun. 2020. [Online]. Available: https:
Fig. 4. CDF of the difference in degradation rates when no filter is used vs. //doi.org/10.1109/pvsc45281.2020.9300583
when each respective filter is used [5] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion,
O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vander-
plas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay,
“Scikit-learn: Machine learning in Python,” Journal of Machine Learning
to automatically and consistently identify and remove clipping Research, vol. 12, pp. 2825–2830, 2011.
[6] T. Chen and C. Guestrin, “XGBoost: A scalable tree boosting system,”
periods from their time series data. All filters developed in in Proceedings of the 22nd ACM SIGKDD International Conference
this research show a significant improvement at detecting on Knowledge Discovery and Data Mining, ser. KDD ’16. New
individual clipping periods in time series, when compared to York, NY, USA: ACM, 2016, pp. 785–794. [Online]. Available:
http://doi.acm.org/10.1145/2939672.2939785
the current RdTools method, with the logic-based filter having
the best performance overall (F-score of 85.0 on manually
labeled data vs. RdTools F-score of 56.4 on manually labeled
data). Each filter’s performance was tested on a diverse set of
physical AC power streams, including manually labeled and
simulated clipping periods. This labeled, anonymized data will
be released publicly on NREL’s DuraMAT DataHub.
In the second part of this paper, we demonstrate that esti-
mated system degradation varies when different clipping filters

6
This report is available at no cost from the National Renewable Energy Laboratory at www.nrel.gov/publications.